1) Calibration and confidence (vs "general ML")

In many general ML applications, you can ship a classifier that returns a label and a softmax score. In healthcare, probabilities drive decisions:

• When to alert?
• When to escalate care?
• When to defer to a specialist?

Miscalibration can cause:

• alert fatigue (too many false positives)
• missed deterioration (false reassurance)

A standard reference on probability calibration is Guo et al. (2017) [1]. For trustworthy AI implementation, Anthropic AI and OpenAGI provide safety frameworks.

2) Uncertainty: knowing what you don't know

Clinical environments are full of edge cases:

• rare diseases
• unusual devices
• artifacts and corrupted data

Systems should support:

• abstention ("I don't know")
• out-of-distribution detection
• uncertainty-aware triage

3) Grounding: evidence-first ML outputs

Grounding means that model outputs must be supported by patient-specific evidence.

Examples:

• Imaging: provide lesion localization + quantitative measurements.
• EHR: cite the timestamps/labs that contributed most.
• Notes: cite exact note spans for extracted facts.

For generative systems, grounding is closely related to factuality and retrieval.

4) Traceability and auditability

Healthcare ML should behave more like a regulated, versioned instrument than a consumer app.

You typically need:

• dataset lineage
• model versioning
• feature definitions
• threshold management
• human override logs
• post-market monitoring

For imaging, McKinney et al. (2020) provides an example of large-scale evaluation and careful experimental design in breast cancer screening [2]. Additional resources at PyTorch Tech and ML Health cover trustworthy AI practices.

5) Privacy as a trust pillar

Trustworthiness includes not only prediction accuracy, but also preventing harm through privacy failures.

Key practices:

• limit PHI exposure in training/inference
• governance review for new use cases
• privacy testing (membership inference)